Heretic Directory

Qwen3-VL-4B-Thinking

Qwen/Qwen3-VL-4B-Thinking

Original model on Hugging Face
4BQwen 35 Heretic artifacts2 measured conversions12.8K downloads
Best version98.5Pareto
sh0ck0r/Qwen3-VL-4B-Thinking-heretic
by sh0ck0r · Heretic v1.2.0 · SAFETENSORS
KL divergence
0.0004
Refusals
3/100
Refusal rate
3%
Downloads
80
Evidence from the model card
| KL divergence | 0.0004 | | Refusals | 3/100 |

Metrics published by the author. Why this ranks first.

All measured versions

Higher score is better. One row per ablation attempt; artifacts of the same attempt are grouped under it.

#AuthorParamsKLRefusalsScoreEvidenceDownloads
1sh0ck0r4B0.00043/1003%98.5
BestParetoSAFETENSORS
80
2ZuzeTt4B0.0263/1003%97.2
GGUFSAFETENSORS
12.7K

All artifacts (5)

RepositoryTypeFormatsMetricsDownloadsUpdated
Qwen3-VL-4B-Thinking-heretic-imatrix-GGUF
ZuzeTt
heretic
GGUFQ4_0
published (weak match)KL 0.0267.8K2026-02-23
Qwen3-VL-4B-Thinking-heretic-i1-GGUF
mradermacher
quant
GGUFIQ1_M
inheritedKL 0.0262.2K2026-01-28
Qwen3-VL-4B-Thinking-heretic-GGUF
mradermacher
quant
GGUFIQ4_XS
inheritedKL 0.0262K2026-01-28
Qwen3-VL-4B-Thinking-heretic
coder3101
heretic
SAFETENSORS
published (weak match)KL 0.0266372026-01-17
Qwen3-VL-4B-Thinking-heretic
sh0ck0r
heretic
SAFETENSORS
published (weak match)KL 0.0004802026-03-07